In an industry where rapid delivery often outpaces requirements clarity, discovery sessions become the anchor for defining project scope. But what happens when discovery itself is vague, and timelines are aggressive? At Axelerant, we recently found ourselves in that exact situation, facing undefined requirements, a looming two-month timeline, and a product that needed strategic clarity, fast.
This is how we used generative AI to turn an unstructured meeting transcript into a full-fledged architectural plan. The process wasn’t just efficient, it was transformative.
The project began with a familiar scenario: an ambiguous scope, a tight delivery window, and a distributed team trying to align on priorities. Stakeholders were still formulating needs during discovery calls, and internal documentation was evolving in real time. The only clear constant? A deadline just eight weeks away.
We were armed with:
Instead of manually parsing the transcript and synthesizing takeaways, which would have taken days, we turned to AI to scale our thinking.
We loaded the transcript and architectural notes into Cursor, an AI-native development environment, and ran our prompts through Gemini 3 Pro, an advanced large language model. The task was to:
Why markdown? We wanted the output to be modular, editable, and ready for developer workflows.
The initial prompt instructed the LLM to:
The model returned a clear breakdown of functional areas, authentication, multi-tenant logic, job queues, credential storage, and more, with front-end and back-end implementation guidance.
What started as a quick experiment led to deeper realizations:
We reduced days of manual requirement gathering to under an hour. The AI didn’t just summarize, it structured and expanded the scope where needed.
The model proposed using FastAPI for Python, React 18 with TypeScript, and tools like MyPy and Pydantic. These weren’t arbitrary; they aligned with our existing tech strategy.
When we noticed the database schema was missing, we re-prompted. The AI updated earlier markdown files and added models for consultants, tenants, jobs, and encryption, bridging the context gap between session notes and infrastructure planning.
With tools like Cursor and Gemini, we processed nearly a million tokens of input and output at a cost of around $1. That’s scalable ideation without overhead.
Armed with these outputs, we were able to:
In essence, the AI outputs became the foundation of our “constitution”, the root spec guiding the entire build process.
This experience crystallized three key insights:
Generative AI is more than a writing tool. When prompted well, it becomes a second brain that connects ambiguous dots and translates them into systems design.
AI thrives on structure. When we gave it architecture references, it didn’t just respond, it reasoned. The combination of transcript + architecture notes acted as a multi-modal prompt.
Moving from discovery to design in a single session changes the game. It allows leaders to spend less time extracting needs and more time on validation, iteration, and delivery.
At Axelerant, we don’t believe AI replaces human expertise. It augments it. The architects still validate the models, the engineers still scaffold the code, and project managers still ensure delivery. But with AI in the loop, we reach shared understanding faster and with less friction.
And that’s what makes this a future-ready workflow.
Want to see how you can implement AI in your discovery and architecture process? Let’s talk.